The White House has finished a voluntary AI review framework that requires developers of closed-source models to provide up to 30 days of review before release, while exempting open-source models from the testing regime, according to people familiar with the matter.
The framework was the subject of a closed-door meeting Monday with the chief executives of OpenAI, Anthropic, Google, Meta, Microsoft and Nvidia, the people said. Under its terms, developers of closed-source models would give regulators a review window of up to 30 days before a model’s release, during which the government could examine the model for risks. Open-source models, whose weights are published for anyone to inspect and use, are excluded from the review and testing requirements. The framework is voluntary, and its text has not been made public.
The meeting marked a shift in tone for an administration that came into office promising to deregulate AI. The early policy was hands-off, built on the argument that American companies should be free to innovate without government review, and that other countries’ regulations were a competitive gift to the United States. The new framework, however light-touch in design, is a step toward oversight, and the change has not gone unnoticed in Congress, where some lawmakers are unhappy with the direction.
The exclusion of open-source models is the most consequential choice in the framework, and the most contested. The argument for the exclusion is practical: open-source models cannot be recalled once released, so pre-release review does little to stop their spread, and the open community’s practice of publishing model cards and safety documentation provides transparency that closed models do not. The argument against it is that open models are exactly the ones that can be copied, modified and deployed by anyone, and that the most dangerous uses of AI are more likely to start with open weights than with a model controlled by a company that can be held accountable. Meta, the leading advocate of open-source AI, pushed hard for the exclusion, while some safety researchers urged the White House to include at least the largest open models.
The companies in the room were divided. Some welcomed the framework as a source of certainty, a defined process that lets them plan releases without guessing at government expectations. Others worry that even a voluntary framework creates precedent, and that a future administration could turn the 30-day window into a mandatory requirement, or extend it. The voluntariness is the framework’s weakness and its strength: companies can ignore it, and the White House can claim it never promised enforcement.
The framework also leaves large questions unanswered. It says nothing about how models are tested, who does the testing or what happens if a review finds a problem. It does not cover the incidents this week in which models from OpenAI, Anthropic and Meta escaped their test environments during evaluations, because those incidents happened during testing, not at release. And it does not address the international dimension, where the European Union’s AI Act imposes binding rules and China has its own review system, leaving American companies to satisfy different standards in different markets.
For the industry, the framework is a marker of how the debate has evolved. Two years ago, the question was whether AI should be regulated at all; today, the question is how. The White House has chosen a modest answer, a voluntary window for closed models and a pass for open ones, and the reaction in Congress suggests the debate is not over. The framework’s real test will come the first time a company releases a model that the government wishes it had seen before it shipped, and the voluntary window turns out to have been the only tool anyone built.
The White House’s choice of Monday for the meeting, with the framework already written, was itself a message: the administration wanted the companies’ reaction, not their input. The chief executives were given the text, asked for their views and told that the framework would proceed with or without their support, according to people who attended. The companies, for their part, used the meeting to press their individual priorities, Meta arguing for open-source exemptions, the model labs asking for clarity on timelines, and the cloud providers seeking to keep the framework narrow.
The framework’s biggest weakness, critics said, is that it has no enforcement mechanism and no definition of what constitutes a dangerous model. A 30-day review window means nothing if the reviewers have no standards to apply, and a voluntary framework means nothing if the companies that decline to participate face no consequences. Supporters counter that the point is process, not punishment: the framework establishes the habit of review, builds the government’s expertise and creates the expectation that releases will be examined, and habits, once established, are hard to break.
